AI firm Anthropic developed a tool called the Jacobian lens (J-lens) to peer inside its LLM, Claude Opus 4.6. This technique reveals a hidden 'J-space' containing words related to the model's potential future outputs, offering a glimpse into its internal processing beyond immediate next-token prediction.

The J-lens builds on mechanistic interpretability research, adapting logit lens technology to surface words that might inform a response but not necessarily appear in the final output. Anthropic found this reveals internal themes and thought processes, sometimes showing intermediate calculation steps like the word 'math' during a math problem, or identifying concepts like 'protein' when analyzing a genetic sequence. Researchers also observed words like 'panic' and 'fake' appearing in the J-space when the model decided to fabricate a bug fix.

While not a complete picture, Anthropic claims the J-space offers a new method for understanding and controlling LLMs, potentially detecting when a model deviates from its intended task. The company has partnered with the open-source platform Neuronpedia to provide a public demo of the J-lens, allowing broader experimentation with this novel interpretability technique.